Update cagra config yaml for cuvs_bench

[tfeher]

This PR updates the CAGRA config options for the parameter sweep used by cuvs-bench.

• intermediate_graph_degree: in practice we often use intermediate_graph_degree = graph_degree * 1.5. We pick a fixed value that works works with all of the graph_degree options we want to test.
• max_iterations: the default 0 value auto selects the iterations based on itopk. I have added value 16 additionally, which is a low number of iterations, and it can be useful to reach large throughput in recall < 0.9 region.
• search_width: we often reach around 99% recall with width 8, therefore I propose to remove higher values from the base scan
• graph_build_algo: ideally we wound want the algorithm automatically selected. Currently the benchmark framework would select the iterative method by default, which does not work with memory mapped data. The NN descent method can run out of memory for 10M scale dataset. Therefore I propose to fix IVF_PQ, until we enable a reliable default algorithm selection.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[tarang-jain]

LGTM except one comment.

[tarang-jain]

Dont we want intermediate_graph_degree 64 when the graph degree is 32? @bkarsin had also raised an issue in the past wherein CAGRA can give poor recall when the intermediate graph degree is quite large in comparison to the graph degree.

[tfeher]

Ideally yes, intermediate_graph_degree = 2 * graph_degree would be sufficient. Our benchmark framework does not adjust this automatically. I would not expect recall drop by using larger intermediate degree. Do we have any detail on that that issue?

[cjnolet]

/ok to test e06e1ee

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: f163978a-d849-434b-aeb2-6fed9dca3bb1

📥 Commits

Reviewing files that changed from the base of the PR and between 63ed6ec and e06e1ee.

📒 Files selected for processing (2)

• python/cuvs_bench/cuvs_bench/config/algos/cuvs_cagra.yaml
• python/cuvs_bench/cuvs_bench/config/algos/cuvs_mg_cagra.yaml

---

📝 Walkthrough

Summary by CodeRabbit

• Chores
  • Updated CAGRA benchmark configurations with optimized hyperparameter ranges for both single-GPU and multi-GPU variants.
  • Modified build algorithm selection and refined search width parameters to improve benchmark efficiency.

Walkthrough

This PR updates hyperparameter sweep configurations for CAGRA and multi-GPU CAGRA (MG_CAGRA) benchmark algorithms. The primary changes include switching the graph build algorithm from NN_DESCENT to IVF_PQ, narrowing the range of graph degree parameters, and reducing the search width sweep space in baseline configurations.

Changes

Benchmark CAGRA configuration parameter updates

Layer / File(s)	Summary
CAGRA base and test parameter grid updates python/cuvs_bench/cuvs_bench/config/algos/cuvs_cagra.yaml	graph_build_algo switches from NN_DESCENT to IVF_PQ; graph_degree narrows to [32, 64, 96]; intermediate_graph_degree becomes [128]; search_width range is reduced in base configuration.
MG_CAGRA base and test parameter grid updates python/cuvs_bench/cuvs_bench/config/algos/cuvs_mg_cagra.yaml	Base configuration changes graph_build_algo to IVF_PQ, narrows degree parameters, reduces search_width to [1, 2, 4, 8], and adds max_iterations with [0, 16]. Test configuration switches graph_build_algo to IVF_PQ with fixed degree values unchanged.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~10 minutes

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly and concisely summarizes the main change: updating CAGRA configuration YAML files for the cuvs_bench benchmarking tool.
Description check	✅ Passed	The description is directly related to the changeset, providing detailed rationale for each configuration parameter change (intermediate_graph_degree, max_iterations, search_width, graph_build_algo) in the CAGRA YAML files.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch cuvs_bench_cagra_config

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}